Decompose, Attend and Detect: A Robust Framework for Time Series Anomaly Detection

Deep learning-based reconstruction frameworks are widely employed in Time Series Anomaly Detection (TSAD) tasks. However, these detectors encounter challenges in generalizing to non-stationary time series data, leading to inaccurate reconstructed outputs and limiting their detection performance. To address this issue, we adopt Time Series Decomposition (TSD) methods to separately reconstruct the seasonal and trend components, as they exhibit clearer patterns for generalization. However, existing TSD methods are sensitive to anomalies, resulting in unreliable decomposed results in TSAD tasks. To bridge this gap, we introduce a novel robust auto-correlation attention mechanism into the detector. Extensive experiments conducted on both simulated and real datasets demonstrate that the proposed method outperforms all 14 competing baseline methods.